Acoustic emission (AE) based condition monitoring is a popular method to inspect a material health condition in many areas[1]. It monitors irreversible structure changes of material through radiation of acoustic waves. Typically, the structure change generates a typical spectrum of acoustic waves starting at 1 kHz, and falling off at several MHz, which can be detected and analyzed via an AE system. Meanwhile, AE signals convey massive information such as rise time, AE energy, AE peak frequency, which are expressive of the failure process. AE based condition monitoring technology exhibits excellent capability of detecting, positioning, and characterizing damage and has been wildly used in monitoring of inner structure changes in bridges, pressure containers, and pipeline systems. In the last decade, the AE technology has been applied to condition monitoring of power modules and shows a promising application scenario in the power electronics field, which help us acquire accurate failure information and making precise perdition of a failure in the power module[2].

In this work, we firstly combined the AE monitoring and a deep learning model - Gated Recurrent Unit (GRU) to monitor and predict the failure of power module as shown in Figure 1[3]. The proposed method integrally analyzes and summarizes the patterns from the time-series data obtained during power cycle tests of power modules and predicts the failures. Specifically, the time-series data consisting of junction temperature, electric power, thermal resistance, and several AE parameters were obtained from AE sensor data and power cycling test. The proposed method (a) automatically extracts common or different time-series patterns among devices and estimates device states, and (b) effectively predicts failures by switching prediction models according to changes in device states.

By time-series analysis of AE sensor data, it was confirmed that the AE parameters obtained from the AE sensor data capture the progression and signs of failures. In addition, by verifying the effectiveness of the time-series pattern detection, it was confirmed that the proposed method can accurately extract the timeseries patterns during normal conditions, the time-series patterns that show signs of failure, and the time-series patterns after failure. Furthermore, by verifying the prediction accuracy and learning efficiency, it was showed that the proposed method is capable of faster learning while retaining the same or better performance compared to existing methods.

Reference

[1] A. Nair and C. Cai, "Acoustic emission monitoring of bridges: Review and case studies," Eng Struct, vol. 32, no. 6, pp. 1704-1714, 2010.

[2] S. Müller, C. Drechsler, U. Heinkel, and C. Herold, "Acoustic emission for state-of-health determination in power modules," in 2016 13th International Multi-Conference on Systems, Signals & Devices (SSD), 2016: IEEE, pp. 468-471.

[3] R. Dey and F. M. Salem, "Gate-variants of gated recurrent unit (GRU) neural networks," in 2017 IEEE 60th international midwest symposium on circuits and systems (MWSCAS), 2017: IEEE, pp. 1597-1600.